Abstract

Biologists have developed experimental methods that can measure the activity of thousands of genes within tens of thousands of individual cells. However, these single-cell measurements are destructive and one cannot follow the high-dimensional dynamics of genes across time in one cell. Similarly, the spatial context of cells is often lost or only known with reduced resolution. Computational methods are therefore used to infer pseudo-temporal orderings of cells or to infer spatial locations. We show how Gaussian processes (GPs) can be used to model temporal and spatial relationships between genes and cells in these datasets. As examples I will show how we use Bayesian GPLVMs with informative priors to infer pseudo-temporal orderings [1] and branching GPs to identify bifurcation points [2]. Gene expression data are often summarised as counts and there may be many zero values in the data due to limited sequencing depth. We therefore recently extended these methods to use negative binomial or zero-inflated negative binomial likelihoods and we show that this can lead to much improved performance over standard Gaussian likelihoods [3]. We are interested in many extensions of these methods and I’ll outline some challenges which I think the probabilistic machine learning community could help us with.

Notes

• [1] Ahmed, S., Rattray, M., & Boukouvalas, A. (2019). GrandPrix: scaling up the Bayesian GPLVM for single-cell data. Bioinformatics, 35(1), 47-54.

• [2] Boukouvalas, A., Hensman, J., & Rattray, M. (2018). BGP: identifying gene-specific branching dynamics from single-cell data with a branching Gaussian process. Genome biology, 19(1), 65.

• [3] BinTayyash, N., Georgaka, S., John, S. T., Ahmed, S., Boukouvalas, A., Hensman, J., & Rattray, M. (2020). Non-parametric modelling of temporal and spatial counts data from RNA-seq experiments. Bioarxiv https://doi.org/10.1101/2020.07.29.227207

• Magnus Rattray is a Professor of Computational & Systems Biology at University of Manchester. His website can be found here.